رئيس مجموعة SK يحذر: نقص رقائق الذاكرة للذكاء الاصطناعي في 2027 سيكون

The global artificial intelligence memory chip shortage is entering its most critical phase, and the head of the world's dominant high-bandwidth memory (HBM) supplier says the worst is yet to come.

SK Group Chairman Chey Tae-won — who controls SK Hynix, the company producing the majority of the world's HBM chips for AI — gave a stark warning in a wide-ranging CNBC interview on August 13, 2026. His message: 2027 will bring the most severe supply-demand imbalance the memory industry has ever seen, prices are already spiking 40–50%, and even the largest chip buildout in history may not be enough to satisfy customer demand.

"Like a war" for memory chips

Chey described the current race to secure AI memory chips as a war-like scramble. "It's like a war," he told CNBC in Seoul. "Everybody wants to buy the memory chips. Without them, you can't build AI computing systems" .

Foreign governments have already started intervening on behalf of their own industries, he noted, and warned that Seoul is likely to face similar pressure. The situation is so dire that Chey characterized it as "utter chaos" in other appearances .

2027: the worst year ahead

Chey projected that 2027 will bring the worst-ever supply-demand imbalance for memory chips. The reason, he explained, is that demand is surging while no memory company has been able to add enough capacity in time .

This timeline is critical. SK Hynix's massive Yongin Semiconductor Cluster — the anchor of its expansion — will begin production in February 2027 . But Chey's own forecast suggests that even that timing may be too late to prevent peak tightness.

Customers have told Chey they expect AI chip demand to increase 60% to 100% next year alone, while supply will barely increase. Chey said those requests amount to "utter chaos" for memory supply .

"Chipflation": memory prices up 40–50%, consumer prices follow

Chey warned that memory prices have risen 40–50%, a phenomenon he termed "chipflation." This is already forcing global IT companies — including Apple — to raise prices on finished consumer products .

He described today's memory prices as "abnormal" and warned that sustained high prices would invite new competitors and geopolitical retaliation. "Prices went up too fast," he said, and he hopes all this buildout will help restore balance .

AI agents fundamentally changed the market

Chey argued that the current shortage is structurally different from past semiconductor cycles. Historically, demand was driven by the number of hardware devices — smartphones, PCs, servers. But in the AI era, a single user deploys more than 10 AI agents, each requiring enormous memory .

He compared today's AI to "a child who has just turned four" — still early in its development but already consuming vastly more memory than expected. As AI matures, memory needs will only compound .

$720 billion expansion: the world's largest memory fab buildout

SK Hynix is undertaking the world's largest memory fab buildout, betting $720 billion to roughly triple production capacity by 2034 .

The cornerstone is the Yongin Semiconductor Cluster in South Korea, where production will begin in February 2027. Chey acknowledged that even South Korea's national plan to double memory production in five years — and SK Hynix's own aggressive buildout — still falls short of what customers need .

"All my customers said, 'Well, that's not enough, man, we need more,'" Chey told CNBC .

Evaluating U.S. fabs: costs double, but "possible"

Chey said SK Hynix is willing to build front-end memory chip production facilities in the United States but voiced frustration over several obstacles. Construction costs are roughly double those in South Korea, and land, power, water, and regulations are all more burdensome .

The company already has a $3.87 billion advanced packaging plant under construction in West Lafayette, Indiana . Chey said building a full memory production plant is also possible if conditions such as power, water, workforce, and supply chain are in place .

Beyond a memory plant, Chey indicated SK Hynix could invest tens of billions of dollars in AI data centers and related technologies, as well as joint ventures .

New business models: LTAs, memory-as-a-service, joint ventures

To mitigate risk with key customers, Chey is exploring long-term agreements (LTAs), "memory-as-a-service" models, and joint ventures with big tech companies .

These new business models reflect the structural change Chey sees: the AI era is creating a permanently higher floor for memory demand, and traditional spot-market pricing may no longer serve either suppliers or customers well.

A shortage that may last until 2030

Chey has previously warned that the memory shortage could persist through the end of the decade. At Computex 2026 in Taipei, he said the global supply gap for high-bandwidth memory will continue through 2030 .

Investors and analysts expect shortages to last through at least 2027, giving memory producers unusual pricing power with some of the world's largest technology customers .

What it means for consumers and the AI industry

The practical impact of Chey's warnings is clear: AI memory chips are becoming a bottleneck for the entire AI infrastructure buildout. Companies building AI computing systems cannot function without them, and the shortage is now pushing costs through the supply chain all the way to consumer prices.

For anyone buying consumer electronics — smartphones, laptops, AI-enabled devices — the "chipflation" Chey describes means higher prices are likely through at least 2027. For the AI industry, it means the race to train and deploy large language models and AI agents will be constrained by memory availability, not just compute.

Chey's message is blunt: the industry needs to plan for a long period of tight memory supply, and even a $720 billion bet may not be enough. "I'm trying to do my best," he said . Whether that's enough remains the central question for the global AI economy.